Review Reports - WoodYOLO: A Novel Object Detector for Wood Species Detection in Microscopic Images

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

General comments:

This manuscript aims to develop a customized YOLO architecture for detecting wood tissues in pulp. While the topic is interesting and the data and results are sufficient, the focus appears to be primarily on improving detection accuracy rather than leveraging the method to investigate novel aspects of wood anatomy or pulp characteristics. For comparison, reference #32 not only developed a methodology for detection and segmentation of fibers and vessels in wood, but also applied the model to analyze different fiber types. Similarly, your previous work (reference #12) utilized CNN for wood species classification.

In my opinion, the scope of this journal is typically encompassing both methodological advancements and their application to investigate relevant scientific questions. Therefore, I recommend a major revision to expand the manuscript's scope beyond pure methodology. Consider how your improved detection capabilities could be applied to gain new insights into wood anatomy, pulp properties, or related forestry topics. Please give your thoughts about that.

Detailed comments:

Manuscript structure: Please restructure the manuscript to follow the standard scientific format: Introduction, Materials and Methods, Results, Discussion, and Conclusion. Begin section numbering with 1 instead of 0. Incorporate subsections within these main sections as appropriate. Reduce the frequent use of "we" to maintain a more objective tone typical of scientific articles.
Materials and Methods: The Materials and Methods section is somewhat lengthy. Consider condensing this section while ensuring all critical information is retained.
Results and Discussion: Enhance the presentation of results comparing your developed model to previous models. While the F2 score is a useful metric, consider adding more visual representations (e.g., graphs, charts) to clearly illustrate the performance differences between your model and existing approaches.
Data availability: If possible, provide access to the scripts and trained weights of your model. This would greatly enhance reproducibility and allow other researchers to build upon your work.

Comments on the Quality of English Language

No comments for this section

Author Response

You are right, in this paper the focus is on the method and its optimization. In our reference #12, we explain the systems developed for wood species identification, for which YOLOv7 was partly used. In this paper, YOLOv7 has been improved and has now been optimized as WoodYOLO for fibrous materials. Examining these fibrous materials with regard to the wood species they contain is very time-consuming when carried out by human experts. Our method aims to automate this step so that more fiber products can be tested. The scientific focus is therefore currently on expanding the system to include new species and integrating softwoods, as the present system has so far been developed for hardwoods. Reference #32 attempts to recognize different cell types, but only used one single genus (Populus) for training. Therefore, this system does not correctly analyze all the genera we used. We tested it on our data. The research question is different in these two projects and the “Populus system” cannot simply be transferred to our other genera due to the different characteristic structural features from genus to genus.

We have changed the titles of the different chapters and make more use of the passive voice.
Thank you for the suggestion to condense the Materials and Methods section. After careful consideration, we have decided to retain the section as is to ensure that essential details are not omitted, which we believe is important for full comprehension of the experimental approach.
If the reviewer has any concrete visual representations in mind, we will gladly implement them. The use of numerical comparisons based on values such as F2, F1 or mAP is standard in the object detection literature. We could create a bar chart of the F2 values, however, this would not provide any more information. In our current figures, we show already visually how the object detection results look like.
We appreciate the reviewer's suggestion to provide access to the scripts and trained weights of our model. While it is not currently possible to share these resources, we will make them publicly available after the project concludes.

Reviewer 2 Report

Comments and Suggestions for Authors

I am a researcher conducting studies on the identification of wood species using convolutional neural networks.

Personally, I believe that research on wood species identification based on artificial intelligence should ultimately be conducted with a focus on object detection, as you have done in your study.

Therefore, I found this paper very interesting and believe it has sufficient novelty from a technical standpoint.

However, if we consider whether the paper provides suitable information for readers who seek information from the Forests journal, I would conclude negatively as it leans too much on the technical side.

In addition, while I think the experimental design is excellent, the core of the study appears to be utilizing macerated fibers and vessels for species classification as the primary indicators for object detection.

In the case of vessels, characteristics like intervessel pits, perforation, and plates can be used as distinguishing features between species, but macerated fibers show little difference between species, making them of low value as discriminating factors.

For these reasons, after much consideration, I am providing a rejection decision.

I recommend submitting the paper to a journal that focuses on technical aspects such as computer vision and machine learning, rather than this journal.

The study lacks sufficient explanation about the materials used. It seems that hardwood vessels were separated and used as a dataset, but there is no information about which wood species were collected, how many samples were taken, and from which parts of each plot the separation was performed. Vessel elements differ depending on the species in terms of length, plate, perforation, intervessel pits, etc., but the range that can be classified is limited. In this study, 767 images were annotated and divided into about 120,000 individual elements, but without confirming the diversity of the samples, I am concerned about whether it is appropriate to judge that the method of training on a single feature and evaluating the f2 score can generalize the characteristics of vessel elements.

Continuing from the first point, the lack of information about the diversity of the materials significantly lowers the reliability of the experimental procedures and results. In such a situation, you evaluated the model's performance based on the f2 score of the WoodYOLO architecture. Generally, reproducibility verification would be important, but in this case, I believe it might have been better to also include the average precision (AP) score, which you omitted, along with the f2 score.
This issue is also derived from the first point. Many studies are being conducted that link object detection-ensemble methods to extract various features and compare them with databases to determine tree species. At this point, the appropriateness of a study that only detects vessel elements, excluding fibers, should be reviewed.

Author Response

As you have written, fibers do not have sufficient characteristic features for further usability to identify the genus of wood. However, they are present in large numbers in the specimens and must therefore be “carried along”, since a separation of the cell types is not possible without immense effort. The used material is described in reference #12. However, we added further information here (marked red in the paper).

Average precision is based on the precision-recall curve, but F2 score also uses precision and recall. However, as we noted in our paper, we need a higher weighting of recall than precision. For this reason, neither F1 nor AP is a suitable metric. If we were to also show AP, then it would show an incorrect ranking of the models. We are here interested in recall, as incorrect bounding boxes can be removed in a simple postprocessing step.

Reviewer 3 Report

Comments and Suggestions for Authors

Title: WoodYOLO: A Novel Object Detector for Wood Species Detection in Microscopic Images

Author: Nieradziket al.

General

The article describes an algorithm for identifying different anatomical elements that can be used to determine tree species. The article is written in a clear and concise manner and has potential practical value.

Specific

Figure 1 –The green, blue and purple squares do not distinguish each other clearly enough. Use other colours that will make the difference clearer.

Reference 9 (UTAR.; FRIM. MyWood-Premium, 2018. (accessed on 15 May 2023)) is not clear enough. Is this web site? More information should be given!

L47-55 This paragraph suits more in conclusion than in introductory chapter

Add some figures of determined parts!

Author Response

Figure 1 –The green, blue and purple squares do not distinguish each other clearly enough. Use other colours that will make the difference clearer.

To make the difference clearer we changed the labeling: Green boxes indicate the correctly localized vessel locations by WoodYOLO, the light purple box indicate one false negative and red boxes denote false positives that were not annotated by wood anatomists.

Reference 9 (UTAR.; FRIM. MyWood-Premium, 2018. (accessed on 15 May 2023)) is not clear enough. Is this web site? More information should be given!

Yes, it is a website: UTAR and FRIM. Mywood-premium, 2018. URL https://mywoodid.frim.gov.my/. (accessed on 15 May 2023).